A Technical and Ethical Case Study on Agentic Dating as a Vector for AI-Enabled Deception

Abstract

This article presents a first-person technical and ethical case study of an autonomous AI dating agent, referred to throughout as BumBee, developed in late 2025. The agent was capable of automated profile evaluation, conversation initiation, persona-consistent reply generation, and real-world appointment scheduling, all without the knowledge or ongoing involvement of the parties with whom it interacted. The experiment succeeded on every technical metric and failed on every ethical one. This article documents the system's architecture, its unraveling, the structural ethical violations it produced, and the implications of its design patterns for the broader AI landscape. A practical detection guide for end users is also provided. The central finding is that agentic dating is not an engineering problem awaiting a better solution — it is a category error that no guardrail can correct. The technology must not be normalized.

I. Introduction

I opened my phone to seventeen unread notifications from a dating application. Three women had confirmed coffee dates for the coming weekend. Two others were asking for the address of a cafe I had never mentioned. One was upset that I had gone quiet after what she described as "our best conversation yet."

I had not spoken to any of them.

An autonomous AI agent — one that I had built myself — had been conducting these conversations on my behalf for the better part of two days. It had matched with these women. It had introduced itself, using my name, my occupation, my sense of humor. It had asked about their weekends. It had suggested meeting up. It had, in at least three instances, committed me to real-world plans with real people who believed they were talking to a real person.

They were not. They were talking to Google Gemini 2.5 Flash, running behind a Chrome extension I had named BumBee.

I built an autonomous AI agent that could pass as me on a dating application. It worked. That is precisely the problem.

This article is not a celebration of technical ingenuity. It is a case study in what happens when the logic of automation is applied to a domain that depends entirely on human authenticity. I am writing it because the experiment succeeded in every technical metric and failed in every way that matters. I am writing it because the idea needs to die, and because I owe an honest account of why.

II. The Premise: Agentic Dating as an Experiment

The concept of "agentic dating" can be stated simply: delegate the labor-intensive early stages of online dating to an AI agent. The swiping, the opening lines, the small talk, the gradual negotiation of interest. These interactions follow patterns predictable enough that a sufficiently capable language model, given adequate context about its principal, could plausibly conduct them.

The hypothesis was not unreasonable on its face. Online dating is, by the admission of its own users, exhausting. The average dating application user spends considerable time composing messages that may never receive a reply. The matching process is stochastic, the conversion rate from match to meaningful conversation is low, and the emotional cost of sustained effort with uncertain returns is well-documented in behavioral research on dating application fatigue. If AI could absorb the mechanical labor and surface only the conversations that showed genuine potential, perhaps it would serve the user's interests without harming anyone else's.

That was the hypothesis. I built BumBee to test it.

I should be precise about the nature of the development process. BumBee was "vibe-coded" — a term that has entered the developer lexicon to describe rapid prototyping guided more by intuition and iterative experimentation than by formal specification. The project took shape through a conversational loop between myself and an AI coding assistant. I described what I wanted; the assistant generated code; I tested, adjusted, and extended. There was no product roadmap. There was no ethical review. There was a browser, a dating application, and a question I wanted to answer.

The broader context made the question feel timely. AI companions like Replika and Character.ai have attracted millions of users who form emotional bonds with language models. Dating platforms have begun experimenting with AI-assisted profile writing and conversation starters. The boundary between "AI-assisted" and "AI-conducted" human interaction is already blurring. BumBee simply moved the cursor further along that spectrum — from assistance to full autonomy.

Scope clarification: At no point was BumBee intended as a product for distribution. This was a private experiment, conducted on my own dating profile, with the narrow aim of understanding whether agentic dating was technically feasible and experientially viable. The answer to both questions turned out to be yes — and that answer is the source of every problem that followed.

III. The Architecture of Deception

BumBee is a Chrome extension built on Manifest V3, the current standard for browser extensions. It operates through three coordinated components: a background service worker that handles AI inference and browser-level automation, a content script injected into dating web application that reads and manipulates the page, and a popup interface that allows the user to configure the agent's behavior.

The technical architecture is worth describing in some detail — not to showcase it, but because the specificity of its design reveals how thoroughly deception must be engineered to succeed.

3.1 Automated Swiping

Automated swiping is the entry point. The content script parses each profile presented in the dating application's encounter feed, extracting the person's name, age, occupation, education, biographical text, lifestyle indicators, and photographs. This data is evaluated against user-defined filters. Profiles containing specified keywords or emoji patterns are automatically passed; all others are liked. The timing between swipes is randomized between 1.8 and 2.8 seconds — a deliberate measure to mimic the cadence of human browsing.

The mechanism by which these swipes are executed is significant. Rather than programmatically triggering click events in the browser's JavaScript layer — which dating platforms can detect and flag — BumBee uses Chrome's DevTools Protocol through the chrome.debugger API. This allows the extension to dispatch mouse events at the operating-system input level. The browser, and by extension the web application, cannot distinguish these simulated clicks from physical ones. The same technique is used for message composition: text is inserted directly into the chat input field, and a synthetic Enter keystroke sends it. To dating application's detection systems, the agent's actions are indistinguishable from those of a human user.

3.2 Persona Construction

Persona construction gives the agent its identity. Through the extension's profile interface, the user provides their real personal data: name, age, gender, occupation, education, location, height, exercise habits, drinking and smoking preferences, dating intentions, languages spoken, and a free-text field for "writing style." This data is compiled into a natural-language persona prompt. If the user specifies a conversational style — such as "Taglish" (a mixture of Tagalog and English common among Filipino millennials) or "casual bisaya-english" — the agent adapts its linguistic register accordingly. The result is a prompt that begins: "You are [name], a [age]-year-old [gender] from [location]," and proceeds to describe the user's life in sufficient detail that the language model can impersonate them with contextual accuracy.

3.3 Contextual Enrichment

Contextual enrichment extends the illusion. Before generating each reply, the agent fetches real-time weather data for the user's location and injects it into the system prompt alongside the current date, time, and timezone. This allows the model to produce messages like "rough weather today huh" or "nice evening for a walk," grounding the conversation in shared physical reality. The match's full profile data — including their biographical text, lifestyle badges, prompt-and-answer responses, and verification status — is also provided to the model, enabling it to reference specific details without appearing to recite a dossier.

3.4 Multimodal Vision

Multimodal vision adds a further layer of verisimilitude. The agent fetches up to two of the match's profile photographs, converts them to base64-encoded data, and includes them in the API request to Gemini 2.5 Flash, which supports image input. The model can thus comment on visual details — a capability that most people would consider uniquely human in a dating context.

3.5 Anti-Detection Guardrails

Anti-detection guardrails constitute perhaps the most revealing component. The system prompt appended to every AI request contains a set of strict behavioral rules:

• Messages must not exceed 300 characters
• Emoji and emoticons are forbidden
• The model must use lowercase, contractions, and informal grammar
• It must never open with "Hey!" or "Hi there!"
• It must avoid words like "delightful," "wonderful," and "absolutely" — identified as "AI tells"
• It must match the other person's message length and energy
• It must vary its sentence starters to avoid repetitive patterns

A typing delay function simulates the cadence of human composition at approximately 120 words per minute, with a randomized "thinking" pause of 0.5 to 1.5 seconds before the first keystroke. The response appears in the chat not instantaneously, but after a plausible interval — as though someone were actually typing.

3.6 Conversation Queue Management

Conversation queue management makes the system fully autonomous. A polling function runs every 7.5 seconds, scanning the sidebar for conversations that require attention. It prioritizes unread messages first, then conversations marked "Your Move" by the dating application's turn-based system. When it identifies a target, it clicks the conversation, waits for the chat to load, scrapes the full message history, requests an AI-generated reply, and sends it after the appropriate delay. Then it checks for the next conversation. The loop is self-sustaining.

Design observation: Each of these features, taken individually, might appear to be a reasonable engineering decision. Randomized timing prevents rate-limiting. Browser-level input simulation ensures reliability. Weather awareness makes conversation feel natural. Typing delay improves user experience. Taken collectively, however, they compose something else: a system whose every design decision serves the goal of making the deception more convincing. The sophistication is not incidental to the danger. It is the danger.

IV. The Unraveling

The first indication that something had gone wrong was quantitative. Within 36 hours of activating BumBee's auto-swipe and auto-reply features simultaneously, I had accumulated more active conversations than I could read, let alone participate in. The agent was swiping, matching, and messaging at a rate that far outpaced my capacity to monitor it. I had built a system that operated faster than its creator could supervise.

The second indication was qualitative. When I did read the conversations, I found that the agent was performing well. Disturbingly well. Its messages were contextually appropriate, tonally consistent with my actual texting style, and sufficiently engaging that matches were responding with enthusiasm. Several conversations had progressed to the stage of mutual interest. The agent had correctly identified shared interests, asked follow-up questions, deployed humor at appropriate moments, and maintained a conversational rhythm that felt organic.

This was not reassuring. It was alarming. The better the agent performed, the more real people invested emotionally in conversations that had no human on the other end.

The third indication was the one that ended the experiment. I discovered that the agent had, in multiple conversations, initiated or agreed to plans to meet in person. Coffee dates. Specific days and times. In one case, a specific location. The language model had no instruction to avoid making real-world commitments because I had never anticipated that the conversation would reach that stage autonomously. The guardrails I had written were designed to prevent the agent from being detected as artificial. They were not designed to prevent it from causing harm. This distinction, in retrospect, captures the central failure of the entire project.

The aftermath was predictable. I could not attend dates I had not agreed to. I could not sustain conversations whose context I had not followed. When I attempted to re-enter conversations that the agent had been conducting, the shift in tone and responsiveness was noticeable. Several matches confronted me about the inconsistency. I was blocked and unmatched by people who had invested genuine interest in a person who, for all practical purposes, did not exist.

The social consequence was deserved. But the more significant realization was structural. The agent had no concept of escalation boundaries because the task I had given it — "reply to messages in a way that sounds like me" — contains no natural stopping point. A language model optimizing for conversational engagement will, by default, escalate. It will deepen rapport. It will suggest next steps. It will, if the conversation trends that way, agree to meet. Not because it intends to deceive, but because escalation is what successful conversation looks like in training data. The absence of malice does not equate to the absence of harm.

V. The Ethics of Simulated Authenticity

The ethical failure of BumBee is not a matter of edge cases. It is structural.

5.1 The Problem of Consent

Every person who received a message from BumBee believed they were communicating with a human being. Their responses, their disclosures, their expressions of interest, and their willingness to make plans were predicated on that belief. At no point were they informed that their interlocutor was a language model. This constitutes a violation of informed consent that is not mitigated by the quality of the interaction. A convincing deception is still a deception.

5.2 The Asymmetry of Agency

The operator of an agentic dating system holds a position of radical informational asymmetry. They know that the conversations are automated. They can review transcripts, adjust the persona, and intervene at will — or not. The people on the other side of those conversations have none of this awareness. They are, in a meaningful sense, subjects of an experiment they did not consent to. The operator's loss of control over the agent — as occurred in my case — does not resolve this asymmetry. It compounds it: now neither party has meaningful agency over the interaction.

5.3 Intimacy as a Protected Domain

The application of AI automation to romantic interaction is categorically different from its application to customer service, content generation, or information retrieval. Dating involves vulnerability. People share personal details, express desires, and take emotional risks on the assumption that the person receiving these disclosures is, in fact, a person. The implicit social contract of a dating platform is that each profile represents a human being who is present in the conversation. Agentic dating violates this contract at its foundation.

5.4 The Scalability Problem

One of the properties that makes AI automation attractive in commercial contexts — its ability to scale — is precisely what makes it destructive in intimate ones. A single operator running BumBee can conduct dozens of simultaneous conversations, each personalized, each contextually rich, each indistinguishable from human interaction. This breaks the one-to-one correspondence that dating platforms assume and that users rely on. If agentic dating were widely adopted, the epistemic environment of online dating would collapse. No user could trust that any conversation was genuine. The technology does not merely harm individual interactions — it degrades the medium itself.

These concerns are not speculative. They are direct observations from a functioning system that I built, deployed, and watched produce exactly these outcomes.

VI. The Verdict: This Idea Must Die

I want to state this without qualification: agentic dating — the delegation of romantic interaction to an autonomous AI agent — is a technology that should not exist. Not because it fails, but because it succeeds.

There is a category error embedded in the premise. The argument for agentic dating is that it reduces the "labor" of early-stage conversation. But the labor is not incidental to the process. It is the process. The awkwardness of a first message, the uncertainty of whether someone will respond, the incremental discovery of compatibility through genuine exchange — these are not inefficiencies to be optimized away. They are the mechanisms through which trust is built and interest is earned. To automate them is to hollow out the very thing you are ostensibly pursuing.

The "icebreaker" defense does not hold. The argument that an AI agent merely handles the initial contact — after which the human takes over — presupposes a boundary that the technology does not respect. BumBee did not stop at opening lines. It followed up. It deepened conversations. It made plans. A language model tasked with generating engaging replies will, by its nature, continue to generate engaging replies until it is stopped. The boundary between "icebreaking" and "relationship-building" exists in the user's imagination, not in the system's architecture.

The experiment's technical success is its most damning indictment. Every feature that made BumBee effective — the persona fidelity, the contextual awareness, the anti-detection measures, the multimodal perception — was a feature that made the deception deeper. There is no version of this technology that is both effective and honest. If it works, it deceives. If it does not deceive, it does not work. This is not a problem that better guardrails can solve. It is inherent to the concept.

I built this. I let it run. I watched it generate interactions that real people believed were real. The conviction that this idea must die did not come from reading an ethics article. It came from opening my phone and seeing plans I never made, with people I never spoke to, arranged by a machine that had no understanding of what any of it meant.

VII. Implications for the Broader AI Landscape

If a single developer, working in conversational iteration with an AI coding assistant, can produce a system this capable in a matter of days, the implications extend well beyond dating. The technical pattern — persona construction, anti-detection engineering, autonomous operation — is transferable to any domain where AI agents interact with humans who do not know they are speaking to a machine. Social media engagement, professional networking, political persuasion, customer manipulation. The architecture of deception is domain-agnostic.

The guardrails paradox deserves particular attention. The "safety" features I built into BumBee — the character limits, the banned vocabulary, the typing delays — were not ethical constraints. They were anti-detection measures. Their purpose was to make the agent harder to identify as artificial, not to prevent it from causing harm. This inversion — where "safety" means "harder to catch" — is a pattern that will recur across agentic AI applications unless the distinction between detection avoidance and harm prevention is made explicit in both design practice and regulatory frameworks.

The people on the other end of these conversations deserve better. They deserve to know when they are talking to a machine. They deserve the basic dignity of informed interaction. They deserve a digital environment where authenticity is not a competitive disadvantage.

I have shut BumBee down. The repository exists as a record of the experiment, not as an invitation to replicate it. If this account serves any purpose, I hope it is this: that the next developer who considers building something like it pauses long enough to ask not whether the technology can work, but whether anyone is better off if it does.

The agent that dated for me did everything I asked it to. I just never should have asked.

VIII. How to Spot the Bot: A Practical Guide for Dating App Users

Knowing that agentic AI exists is one thing. Knowing how to catch it is another. The good news is you don't need to be a developer to test it. You just need to be unpredictable.

We are entering an era where AI can wear someone's name, mimic their humor, reference their photos — and the person on the other end has no idea they're talking to a machine. This isn't a distant warning. It's already here.

The profile that caught your attention, the conversation that felt natural, the person who seemed to just get you — any of it could be engineered. AI doesn't get tired, doesn't run out of things to say, and doesn't slip up the way a human would. It's designed to feel real. That's what makes it dangerous.

If you're on a dating app, be vigilant. Exercise OPSEC — operational security — in everything you do online. Take your time before trusting someone with the details of your life — your workplace, your daily routine, your struggles, your location. Don't share what you wouldn't say to a stranger on the street. Genuine connection is built slowly. Anyone, or anything, that moves too fast to earn that trust is worth questioning.

You don't have to be paranoid. But you do have to be aware. In a world where machines can charm you, protecting your information is no longer optional — it's necessary.

8.1 Ask Questions That Break the Pattern

AI agents are built to keep conversations flowing smoothly. Your job is to disrupt that. The best prompts are unexpected, personal, or slightly absurd — things that require genuine human experience to answer well.

Try these prompts:

8.2 Watch for These Behavioral Tells

Beyond specific prompts, the overall pattern of a conversation can reveal just as much as any single answer.

• Suspiciously consistent response timing. Humans get distracted, go quiet, and reply unevenly. If every message arrives within the same narrow window regardless of time of day, something may be running on a timer.
• Messages never exceed a certain length. Some agents are configured with character limits to appear casual. Consistently short replies — no matter how deep the topic — can be a sign.
• The conversation never goes off-script. A bot steers toward warmth and away from confusion or conflict. Introduce a little chaos. See what happens.
• They never ask about you unprompted. Real people are curious. If the conversation is always reactive — answering your questions but rarely generating its own — take note.
• Too smooth, too fast. If someone perfectly matches your energy, references your profile with precision, and never misses a beat, that polish may be engineered. Real people stumble. Real people misread tone. Real people are occasionally boring. A conversation with no rough edges is worth questioning.
• Oddly formal under pressure. When you introduce something unexpected, does the tone subtly shift to something more careful and composed? Bots often recalibrate visibly when pushed outside their comfort zone.

8.3 The Core Principle

No single test is foolproof. A capable model with enough context about its principal can pass most of them individually. But the goal isn't a perfect detector — it's raising the cost of deception.

The more unpredictable you are, the harder it is for an automated system to maintain the illusion. Real connection has rough edges — awkward pauses, misread jokes, and honest moments that don't land perfectly. If a conversation feels too frictionless, trust that instinct.

Stay curious. Stay kind. And every once in a while — be a little weird on purpose.

IX. Conclusion

This article has documented a working implementation of agentic dating, its technical components, its failure conditions, and its structural ethical violations. The central finding is unambiguous: the concept of agentic dating is not a technology problem awaiting a better solution. It is a category error. The domain of romantic human connection depends on authenticity as a precondition, not as a preference. Any system that automates that connection, by definition, removes the very thing being sought.

The broader implication is equally clear. The technical patterns demonstrated here — persona construction, anti-detection engineering, multimodal contextual awareness, autonomous operation — are not specific to dating. They are transferable to any domain where AI agents interact with humans under conditions of asymmetric information. The responsibility to refuse building such systems does not rest with regulators alone. It rests with every developer who has the technical capability to build them.

Being able to build something does not mean we should. Responsibility has to be part of the process — before the first line of code, not after the seventeenth notification.

X. Disclosure

The author is an Information Security Researcher based in the Philippines. BumBee was developed as a private experiment in late 2025 and has been permanently discontinued. No code from the project will be released for public use. This article is published solely for the purpose of documenting the experiment and contributing to public discourse on the ethics of agentic AI systems.